agent-loops（Agent 循环）

Agent Loops

Enable LLMs to reason, plan, and take autonomous actions.

ReAct Pattern (Reasoning + Acting)

REACT_PROMPT = """You are an agent that reasons step by step.

For each step, respond with:
Thought: [your reasoning about what to do next]
Action: [tool_name(arg1, arg2)]
Observation: [you'll see the result here]

When you have the final answer:
Thought: I now have enough information
Final Answer: [your response]

Available tools: {tools}

Question: {question}
"""

async def react_loop(question: str, tools: dict, max_steps: int = 10) -> str:
    """Execute ReAct reasoning loop."""
    history = REACT_PROMPT.format(tools=list(tools.keys()), question=question)

    for step in range(max_steps):
        response = await llm.chat([{"role": "user", "content": history}])
        history += response.content

        # Check for final answer
        if "Final Answer:" in response.content:
            return response.content.split("Final Answer:")[-1].strip()

        # Extract and execute action
        if "Action:" in response.content:
            action = parse_action(response.content)
            result = await tools[action.name](*action.args)
            history += f"\nObservation: {result}\n"

    return "Max steps reached without answer"

Plan-and-Execute Pattern

async def plan_and_execute(goal: str) -> str:
    """Create plan first, then execute steps."""
    # 1. Generate plan
    plan = await llm.chat([{
        "role": "user",
        "content": f"Create a step-by-step plan to: {goal}\n\nFormat as numbered list."
    }])

    steps = parse_plan(plan.content)
    results = []

    # 2. Execute each step
    for i, step in enumerate(steps):
        result = await execute_step(step, context=results)
        results.append({"step": step, "result": result})

        # 3. Check if replanning needed
        if should_replan(results):
            return await plan_and_execute(
                f"{goal}\n\nProgress so far: {results}"
            )

    # 4. Synthesize final answer
    return await synthesize(goal, results)

Self-Correction Loop

async def self_correcting_agent(task: str, max_retries: int = 3) -> str:
    """Agent that validates and corrects its own output."""
    for attempt in range(max_retries):
        # Generate response
        response = await llm.chat([{
            "role": "user",
            "content": task
        }])

        # Self-validate
        validation = await llm.chat([{
            "role": "user",
            "content": f"""Validate this response for the task: {task}

Response: {response.content}

Check for:
1. Correctness - Is it factually accurate?
2. Completeness - Does it fully answer the task?
3. Format - Is it properly formatted?

If valid, respond: VALID
If invalid, respond: INVALID: [what's wrong and how to fix]"""
        }])

        if "VALID" in validation.content:
            return response.content

        # Correct based on feedback
        task = f"{task}\n\nPrevious attempt had issues: {validation.content}"

    return response.content  # Return best attempt

Memory Management

class AgentMemory:
    """Sliding window memory for agents."""

    def __init__(self, max_messages: int = 20):
        self.messages = []
        self.max_messages = max_messages
        self.summary = ""

    def add(self, role: str, content: str):
        self.messages.append({"role": role, "content": content})

        # Summarize old messages when window full
        if len(self.messages) > self.max_messages:
            self._compress()

    def _compress(self):
        """Summarize oldest messages."""
        old = self.messages[:10]
        self.messages = self.messages[10:]

        # Async summarize would be better
        summary = summarize(old)
        self.summary = f"{self.summary}\n{summary}"

    def get_context(self) -> list:
        """Get messages with summary prefix."""
        context = []
        if self.summary:
            context.append({
                "role": "system",
                "content": f"Previous context summary: {self.summary}"
            })
        return context + self.messages

Key Decisions

Common Mistakes

• No step limit (infinite loops)
• No memory management (context overflow)
• No error recovery (crashes on tool failure)
• Over-complex prompts (agent gets confused)

Related Skills

• function-calling - Tool definitions and execution
• multi-agent-orchestration - Coordinating multiple agents
• langgraph-workflows - Stateful agent graphs

Capability Details

react-loop

Keywords: react, reason, act, observe, loop Solves:

• Implement ReAct pattern
• Create reasoning loops
• Build iterative agents

tool-use

Keywords: tool, function, call, execution Solves:

• Implement tool calling
• Execute functions from LLM
• Parse tool responses

workflow-template

Keywords: template, workflow, agent, typescript Solves:

• Agent workflow template
• TypeScript implementation
• Copy-paste starter